Do AI-generated ads work? What the tests show, beyond the TikTok “systems”

By Andrej Ruckij · November 1, 2026

TL;DR: For some jobs, yes. AI-made ads hold up best on simple, product-led briefs. In a Google Display field test, fully AI-made ads had up to a 19% higher click-through rate than human ones. They fall short on story and emotion: in an Ipsos and Syracuse University test of ads from 10 brands, the human versions beat their AI-made counterparts on predicted sales impact. My read is that AI fits static product ads, native formats like text-message screenshots, offers, and localised versions of ads that already work. The real gain is cheap, genuinely different concepts to test, which is what Meta’s and TikTok’s systems now ask for. That only pays off if your budget can test them properly, and most small budgets can test far fewer ads than the TikTok sellers suggest. Label AI ads where platforms require it; an “AI-generated” label cut click-through by about 31.5% in one test, so the ads have to work on their own merits.

Primores builds performance creatives for paid social, including AI-made variants: see our performance creative work.

What the TikTok systems actually do

I watched the four videos behind this piece. None of them shows a prompt, a spend figure or a result you can check, but they describe the workflow clearly:

Sam Butler’s static ad system, which he says he charges $2,000 to build: one product goes in and 40 static ads come out, in formats like claymation, text-message screenshots and fake tweets. You tick the ones you like and the next batch builds on them; the ones you reject are excluded.
Jimmy Reilly’s three-prompt routine, run weekly with Claude: find the best-selling competitors in a niche and their most viral content, extract the common structures into a template, then write your own posts from that template. He says it brought 27,000 leads to a digital product, through organic Reels with no ad spend. That’s not paid ads, and it can’t be checked.
MoneyPrinterTurbo, an open-source project (about 137,000 GitHub stars, MIT licence): topic in, finished short video out, with script, stock footage, voiceover, subtitles and music, and optional auto-posting. The creator promoting it passes on that some developers find setup “clunky if you’re not comfortable in a terminal”.
DropMagic: a dropshipping store built from a supplier link, with generated photos, reviews and a countdown timer, then “make advertisements just like this”. A reviewer estimates an hour or two of copy clean-up per store.

The useful part is the loop: research competitors, turn what works into angles, generate many variants, keep what wins. The part I’d skip is anything that publishes without a person checking it.

What the tests show

The evidence points in two directions, and the split is the useful part.

Where AI ads matched or beat human ones:

Google Display field test (NYU and Emory): fully AI-generated ads had up to a 19% higher click-through rate than human ads, a best case rather than an average. Human ads that were only edited with AI didn’t improve. The same researchers also tested labels: calling an ad “AI-generated” cut its click-through rate by about 31.5% compared with unlabelled human ads.
Taboola ad data (500 million impressions; researchers from Columbia, Harvard, TU Munich and Carnegie Mellon): AI ads had a 0.76% click-through rate against 0.65% for human ones, but after the study’s strictest statistical controls the two were comparable. AI ads that looked artificial did worse, and a large, clear human face mattered most for making an ad feel human and trustworthy.
Meta’s own numbers (the platform reporting on its own tools): campaigns using its Advantage+ creative image generation saw an 11% higher click-through rate and 7.6% higher conversion rate than campaigns without it.

Where they fell short:

Ipsos and Syracuse University (May 2026): they paired human-made ads from 10 brands, including Cheerios, H&M, Visa and TurboTax, with AI-made versions, 20 ads in all, and tested them on 3,000 US consumers. On Ipsos’s measure of predicted short-term sales impact, the human ads came out 11 points above benchmark and the AI versions 5 points below. As Syracuse’s write-up of the study puts it, “AI performed best when the brief was straightforward and product-driven, but struggled when the creative challenge called for storytelling, emotion or a genuine point of view.”

One practitioner case with numbers: Anthony Morcillo described on RevenueCat’s blog how an app called Mojo tested AI video ads. Generic AI avatars “failed almost immediately”. A custom avatar of the product manager reached 87% of the human ad’s conversion rate at $20 a video against $500 for a creator, which worked out to a 31% lower cost per acquisition. In Brazil, an AI-dubbed version of a proven human ad had 40% lower acquisition cost than ads from native creators. A 0.2-second lip-sync delay on one key word cut click-through by 68%. An AI avatar dubbed into another language worked in Brazil and Spanish-speaking markets and failed in France and Germany. It’s one app and self-reported, but it shows how narrow the winning zone can be.

Where AI helps and where it doesn’t

Putting those together:

Good fit: product-led static ads; native formats such as text messages, notes-app screenshots, social-post mockups, letters and offer banners; translating and dubbing ads that already work; quick variants of a proven concept for new audiences.
Poor fit: ads that rely on story, emotion or a point of view; generic stock avatars; anything polished enough to read as an ad in the first second; any ad that looks AI-made.

The native formats are worth a note. In Motion’s 2026 benchmark of more than 550,000 Facebook and Instagram ads (vendor data, weighted to the fourth quarter), formats like text messages, social-post mockups and ChatGPT screenshots were rare but got more than their share of budget: text-message ads were 0.55% of ads and received 1.2 times their share of spend. That’s correlation from one season, not proof, but these are exactly the formats AI makes cheaply.

Why genuinely different ads matter more than volume

Meta’s ad system now does more of the audience-finding itself, and it uses the creative to decide who sees what. Meta’s guidance is direct: if your ads are too similar, the system “may not recognize it as different and could deliver it to the same sets of audiences”, and you may see “Creative limited” or “Creative fatigue” in the Delivery column of Ads Manager.

What Meta counts as different is a different theme, message or visual for a different kind of buyer: deals and value for younger shoppers, reviews and demonstrations for older ones, the product shown in a new situation or framed as a conversation. For Advantage+ shopping campaigns it suggests at least 20 diversified ads and 20-30% of budget for testing new ones. Jon Loomer, who writes about Meta ads, puts the catch simply: “Unless they’re distinctly different enough, any variation is unlikely to make a meaningful impact on performance.” TikTok’s own guidance is 3-5 unique creatives per ad group.

So 40 ads from one prompt are only worth having if they’re 40 different ideas. Forty colour swaps of one ad are likely to be treated as close to one.

How to build the system

Research in Meta’s Ad Library, filtered to an EU country. For ads shown in the EU you can see the start date, whether there are several versions, and reach by age and gender; outside the EU you get far less. Collect your competitors’ current ads and write down the angle and format of each.
Write angles, not prompts. List 5-10 distinct reasons someone would buy (a problem solved, a comparison, an offer, a review, a use case) and match each to a buyer type. This is the part AI can help with but shouldn’t own.
Generate per angle, in formats AI handles well. For each angle, make a few native-format statics or a localised version of a proven video. Keep real-looking faces large where people appear.
Check every ad before it goes out. Wrong product details, garbled text, odd hands and anything that looks AI-made get cut. Meta’s AI terms make the advertiser, not Meta, responsible for what the tools produce.
Launch in one place. A few meaningfully different concepts per ad set (Loomer’s own set-up is two flexible-format ads with ten creatives each) rather than many near-copies across many ad sets.
Refresh weekly with new concepts, and watch the Delivery column for fatigue.

Tools and what they cost

Option	What you get	Price (as listed, November 2026)
Image models through an API (Google’s Nano Banana family)	Raw images; you build the workflow	about $0.04-0.13 per image at standard rates, about half in batch; 100 drafts cost roughly $4-14
Creatify	AI avatar video ads, plus a competitor ad tracker and ad cloning on Pro and up	from about $10-15 a month (about 11 short avatar videos) to $136-195 (about 300)
AdCreative.ai	Generated ad sets with performance predictions	from about $20 to $999 a month, and its pricing page shows conflicting figures; its “up to 14x more conversions” is a marketing claim
Arcads	AI avatar UGC videos	no public price
MoneyPrinterTurbo	Free, open-source short-video pipeline	free, plus your time; a review notes stock clips can be “technically related to a keyword but wrong for the sentence”

Generating is now cheap. What limits you is the ad spend needed to test what you generate.

How many ads your budget can actually test

A common rule among media buyers is to give each new concept 2-3 times your target cost per acquisition before judging it (a guide from Spark UGC, which sells ad creative, suggests a $100-150 minimum). At a $40 target CPA the $100-150 floor applies, so a $1,000 monthly testing budget covers about 7-10 concepts a month, not 40.

Two numbers explain why that’s tight:

Most ads don’t win. In Motion’s benchmark, about 5% of ads became winners (spending at least 10 times the account’s median ad), and roughly half got no meaningful spend at all. Launching more ads produced more winners, but “it doesn’t make the average ad better.”
Small numbers are noisy. My rough arithmetic: an ad that truly converts at your target CPA will still show zero conversions after 3 times CPA in spend about 5% of the time, so a strict kill rule throws out about 1 good ad in 20. Telling an ad that’s twice as good from an average one needs roughly 20 or more conversions per ad.

What that means in practice: judge new ads first on cheap early signals, such as hook rate (the share of viewers who watch past the first seconds) and click-through, and judge cost per purchase only on the ads that survive. Meta’s learning phase settles after about 50 optimisation events, such as purchases, in 7 days in an ad set, and TikTok’s after about 50 conversions, which is another reason to keep tests in one place.

The rules

Meta: ads “created or significantly edited” with AI get an “AI info” label; under the 2025 rules, for photorealistic AI people it sat next to “Sponsored”, and since June 2026 the label also shows under “About this ad” in the ad’s menu. Meta says it will automatically detect ads made with third-party AI tools through industry-standard signals, and its terms say you must not “misrepresent Output as human-generated” or remove watermarks or authenticity metadata.
TikTok: fully AI-generated images, video or audio need TikTok’s AI-generated content label or your own clear disclaimer, or “your ad will be rejected or restricted.” Using AI voice cloning to make someone say something they didn’t counts as significant AI editing that must be disclosed. Fake endorsements and misuse of public figures are banned, and TikTok says it will remove ads that use a private person’s likeness without permission.
YouTube: mass-produced, repetitive videos can’t earn money, and since July 2025 YouTube calls them “inauthentic content”. That’s a rule for creators, but it’s a warning about unattended pipelines.
New York: ads with a synthetic person who isn’t a real performer must disclose it, since June 9, 2026, with fines of $1,000 and then $5,000.

The label costs clicks, as the NYU and Emory study showed. That’s a reason to make ads that work despite it, not a reason to hide it.

Who this is for

For: brands and small teams running Meta or TikTok ads who make their own creative or brief an agency, with a testing budget of at least a few hundred a month. Not for: ads built on a recognisable real person, regulated claims (health, finance), where every variant needs review anyway, or brand campaigns where the story is the point.

What I’d do this week

Pick one product. Write five genuinely different angles for it, make two native-format static ads per angle with AI, check each one, and run them in one ad set against your current best ad with a fixed budget of about 2-3 times your target CPA per angle. Judge on hook rate and click-through first, and cost per purchase only on the survivors.

More on making ads that test well: glossary/advantage-plus and Should your clothing brand use AI models?

Sources

More of this every day on the channel: https://t.me/aidaily_business